Big Tech just proved the AI spending bill is starting to pay off. And they raised the bill.

Yesterday in AI | Thursday, April 30, 2026

Big Tech just proved the AI spending bill is starting to pay off. And they raised the bill.

Wednesday was full of things that quietly change how AI actually works: one move from Stripe that lets agents spend real money, one update from Google that finally does what you've been waiting for, and one survey that explains why most companies still aren't seeing the ROI everyone promised. Plus: Musk on the stand, Anthropic's trillion-dollar moment, and a hotel booking feature that says more about AI development speed than it seems.

The number that stuck with us: only 16%. Listen to find out why.

Remember to subscribe, rate, and share this podcast if you like it!

Transcript

SPEAKER_00 0:03

Yesterday in AI. Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of artificial intelligence in 10 minutes or less. I'm Mike Robinson. It's Thursday, April 30th, and Big Tech just told us what all that AI spending actually looks like on a balance sheet. AI agents got their first credit card, and Jim and I learned to build your files on the fly. Let's get into it. Before we get into Wednesday's news, a quick catch up on a couple of Tuesday stories we haven't covered yet. The Musk vs. OpenAI trial officially kicked off in federal court in Oakland. Elon Musk took the stand himself on day one. His lawyer opened with The Defendants in this case stole a charity, arguing that Altman and co-founder Greg Brockman raised hundreds of millions under a nonprofit charter, then converted OpenAI into a for-profit company now worth hundreds of billions. OpenAI's lawyer countered that Musk sued because he didn't get his way and that he tried to seize control in 2018, was turned down, walked out, and only filed this lawsuit after the company became one of the most valuable startups in history. The trial runs about three weeks. The judge, not a jury, will decide whether OpenAI should be forced back into nonprofit form. Musk wants Altman and Brockman removed from the board and the IPO blocked. His lawyers concede a full sweep is unlikely. Day two continued Wednesday. There's a lot more coming. Also Tuesday, Anthropic crossed a$1 trillion valuation, officially becoming the most valuable AI company on Earth. That happened on the same day The Atlantic published a piece about OpenAI titled Anthropic's Little Brother. The timing was not subtle. Now to Wednesday. Microsoft, Alphabet, Amazon, and Meta all reported Q1 earnings Tuesday evening, and the results landed with a clear theme. The AI spending is holding, and early revenue signals are justifying more of it. Microsoft reported cloud revenue up 20% year over year, with Azure growing 35%, significantly ahead of analyst expectations. Microsoft said AI services contributed 16% points of Azure's growth, up from 13 last quarter. That's the kind of solid attribution investors have been waiting for. Alphabet reported Google Cloud Revenue up 28%, also ahead of estimates, with CEO Sundar Bachai saying AI-powered products are now driving meaningful revenue across search, workplace, and cloud. Meta's results were similarly strong, with revenue up 16% and the company raising its full-year CapEx guidance to between 64 and 72 billion, up from an earlier estimate of 60 to 65 billion. Meta said the increase is primarily driven by accelerating our AI infrastructure investments. Amazon reports separately, but AWS revenue has been expected to show similar AI-driven growth. The bottom line, going into this earnings cycle, the question was whether the massive infrastructure spending would start showing up in actual revenue. Wednesday's coverage suggests it is. That's a meaningful signal for everyone trying to read where enterprise AI adoption sits in the cycle. It's still early, but the money is starting to flow back. One of the more genuinely new developments Wednesday was out of Stripe. AI agents can now spend money on your behalf. Stripe announced link wallet integration that gives agents a one-time use virtual card, locked credentials, and per-purchase approval gates. Setup is a single NPM install on a configuration file. Until now, agents could plan, draft, search, and call APIs. Spending actual money required a human to step in. That's the step Stripe just removed. The guardrails are real, spending limits, one-time cards, explicit approvals. But the conceptual shift matters. An agent that can spend money is a different kind of tool than one that only makes recommendations. It's the difference between an AI that finds the best flight and one that actually books it. How those guardrails hold up at scale will be one of the more interesting things to watch in the agent space over the next six months. Wednesday also brought something from Google that's been a missing piece in AI assistance for a while. Gemini can actually generate actual files inside chat. Type a prompt, get a PDF, Word doc, Excel Sheet, or Google Docs, Sheets, or Slides export. No copy pasting. No opening a separate app. It's live for every Gemini app user globally right now. This sounds simple, but it closes a significant gap. Some AI assistants will write content for you but can't hand you a finished file. You still have to do the transfer yourself. Gemini is doing the last step. For anyone using AI and document heavy workflows, that's worth testing today. From files to hotels, here's a story worth paying attention to if you're watching how Agentic development is changing software companies. Uber announced Wednesday that US customers can now book hotels directly inside the Uber app, with access to more than 700,000 hotels through an Expedia partnership. Uber One members get discounts on 10,000 hotels and earn Uber credits on bookings. The AI angle is in the build speed. Uber's CTO said Agentic Coding Tools, specifically Cursor, helped compress development timelines for features like hotel booking roughly in half versus historical expectations. So a capability that might have taken 18 months shipped in closer to 9. This is a real-world data point on what AI-assisted development actually does to product velocity at a company with thousands of engineers. Uber isn't a startup. They have engineering process, compliance requirements, platform dependencies. And they still cut the build time in half. If that holds across the industry, competitive motes that depend on engineering throughput start looking a lot thinner. A different kind of story Wednesday, but an important one if you work in content or compliance. Utah enacted a new deepfake takedown law targeting non-consensual AI-generated explicit images. Under the law, platforms operating in Utah must provide a way for individuals to report AI-generated intimate images made without consent. Once reported, platforms have 48 hours to remove the image and make reasonable efforts to pull identical copies. Platforms also have to disclose provenance metadata showing an image's history and how it was altered. The 48-hour removal window is aggressive by current content moderation standards. The identical copy requirement pushes platforms toward more robust detection tooling. For companies running content platforms, state laws like this tend to become de facto national standards once enforcement starts. Easier to build one compliance system than 50. Last story a new Harvard Business Review survey out Wednesday put hard numbers on a problem most enterprise teams already feel. AI is in production at a lot of companies, but the value isn't showing up at the same rate. 59% of organizations said they have AI in production. Only 16% said they've seen a high degree of measurable business value from it. 36% said the value has been slight. The biggest gap, only 18% said AI is primarily integrated within workflows. 34% said they're using it as standalone tools alongside existing processes. That distinction is the one that matters. A standalone AI tool is a productivity add-on. AI integrated into a workflow is a process change. The survey found the top barriers are siloed data, lack of system integration, and missing AI talent. None of those are surprising. All are fixable. But they take longer to fix than it takes to buy a tool subscription. The companies that figure out integration first are the ones that will actually see the return on investment. One more thing.